{"id":8625,"date":"2026-05-30T16:24:30","date_gmt":"2026-05-30T08:24:30","guid":{"rendered":"https:\/\/infernews.com\/blog\/contrastive-probing\/"},"modified":"2026-05-30T16:24:30","modified_gmt":"2026-05-30T08:24:30","slug":"contrastive-probing","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/contrastive-probing\/","title":{"rendered":"contrastive-probing\uff1a\u62c6\u89e3 VLM \u7a7a\u9593\u5224\u65b7\u504f\u5dee\u7684\u8f15\u91cf\u8a3a\u65b7\u9805\u76ee"},"content":{"rendered":"<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/05\/pasted-e3ece1535bf7.jpg\" alt=\"Repository image for cheolhong0916\/contrastive-probing\"><\/figure>\n<p>contrastive-probing \u662f\u4e00\u500b\u7528\u4f86\u6aa2\u67e5 Vision-Language Models\uff08VLMs\uff09\u5167\u90e8\u7a7a\u9593\u8868\u793a\u7684\u8f15\u91cf\u9805\u76ee\uff0c\u7126\u9ede\u4e0d\u662f\u6a21\u578b\u7b54\u5c0d\u5e7e\u591a\u984c\uff0c\u800c\u662f\u5b83\u8166\u5167\u5982\u4f55\u5206\u958b left \/ right\u3001above \/ below\u3001far \/ close\u3002\u5b83\u6cbf\u7528\u8ad6\u6587\u300aWhy Far Looks Up: Probing Spatial Representation in Vision-Language Models\u300b\u7684 contrastive probing \u65b9\u6cd5\uff0c\u900f\u904e\u4ea4\u63db\u554f\u984c\u4e2d\u7684\u5169\u500b\u7269\u4ef6\uff0c\u518d\u6bd4\u8f03 hidden states \u5dee\u7570\uff0c\u62bd\u51fa \u0394 vectors \u4f5c\u5206\u6790\u3002<\/p>\n<p>\u4f7f\u7528\u6642\uff0c\u505a\u6cd5\u662f\u628a\u4e00\u5f35\u5716\u7247\u914d\u4e0a\u4e00\u689d\u539f\u59cb\u7a7a\u9593\u554f\u984c\uff0c\u518d\u751f\u6210\u4e00\u689d\u4ea4\u63db obj1 \u2194 obj2 \u7684\u5c0d\u7167\u554f\u984c\uff0c\u4e4b\u5f8c\u5c0d\u540c\u4e00\u500b VLM \u8dd1\u5169\u6b21 forward\uff0c\u4e26\u5728\u6bcf\u5c64 transformer \u64f7\u53d6\u6700\u5f8c token \u7684\u8868\u793a\u3002\u9019\u500b\u6d41\u7a0b\u53ef\u914d\u5408 \ud83e\udd17 transformers \u8f09\u5165\u7684\u6a21\u578b\uff0c\u7136\u5f8c\u8f38\u51fa Axis Coherence\u30016\u00d76 \u0394-similarity heatmap\u30012D\/3D PCA \u8996\u89ba\u5316\uff0c\u4ee5\u53ca Vertical\u2013Distance Entanglement Index\uff08VD-EI\uff09\u7b49\u7d50\u679c\u3002<\/p>\n<p>\u9019\u500b\u9805\u76ee\u89e3\u6c7a\u7684\u6838\u5fc3\u554f\u984c\uff0c\u662f benchmark accuracy \u5f80\u5f80\u53ea\u544a\u8a34\u4f60\u6a21\u578b\u6709\u5187\u7b54\u4e2d\uff0c\u537b\u672a\u5fc5\u63ed\u793a\u5b83\u662f\u5426\u7528\u5c0d\u4e86\u7a7a\u9593\u7dda\u7d22\u3002\u8ad6\u6587\u8207\u9644\u5e36\u8aaa\u660e\u6307\u51fa\uff0c\u591a\u500b\u6a21\u578b\u5bb6\u65cf\u90fd\u51fa\u73fe vertical-distance entanglement\uff0c\u4e5f\u5c31\u662f\u628a\u756b\u9762\u8f03\u9ad8\u7684\u4f4d\u7f6e\u8aa4\u7576\u6210\u8f03\u9060\uff0c\u53cd\u6620\u81ea\u7136\u7167\u7247\u5e38\u898b\u7684 perspective heuristic\u300chigher in the image \u21d2 farther away\u300d\u3002<\/p>\n<ul>\n<li>\u7528\u6700\u5c11\u5c0d\u7167\u8a2d\u8a08\u89c0\u5bdf\u8868\u793a\u5c64\uff0c\u800c\u4e0d\u53ea\u770b\u7b54\u984c\u5206\u6578<\/li>\n<li>\u53ef\u6bd4\u8f03\u4e0d\u540c layer \u7684\u7a7a\u9593\u8ef8\u662f\u5426\u6e05\u6670\u5206\u96e2<\/li>\n<li>\u80fd\u767c\u73fe vertical \u8207 distance \u662f\u5426\u7cfe\u7e8f\uff0c\u5e6b\u52a9\u627e\u51fa\u504f\u5dee\u4f86\u6e90<\/li>\n<li>\u9069\u5408\u5206\u6790 EmbSpatial-Bench\u3001SpatialTunnel \u9019\u985e\u7a7a\u9593\u63a8\u7406\u8cc7\u6599<\/li>\n<\/ul>\n<p>\u5c0d\u7814\u7a76\u8005\u3001\u6a21\u578b\u5206\u6790\u4eba\u54e1\uff0c\u6216\u8005\u8981\u6aa2\u67e5 multimodal assistant\u3001robotics\u3001embodied agents \u80cc\u5f8c\u7a7a\u9593\u63a8\u7406\u53ef\u9760\u6027\u7684\u4eba\uff0c\u9019\u500b\u9805\u76ee\u5c24\u5176\u6709\u7528\u3002\u73fe\u6709\u8cc7\u6599\u986f\u793a\uff0c\u5c31\u7b97 benchmark \u5206\u6578\u76f8\u8fd1\uff0c\u4e0d\u540c VLM \u7684\u5167\u90e8\u8868\u793a\u4e5f\u53ef\u80fd\u5dee\u5f88\u9060\uff0c\u800c\u7a7a\u9593\u8ef8\u5206\u5f97\u8f03\u6e05\u695a\u7684\u6a21\u578b\uff0c\u901a\u5e38\u5728\u4e0d\u540c\u6e2c\u8a66\u4e0a\u7684\u7a69\u5065\u6027\u6703\u8f03\u597d\u3002<\/p>\n<p>\u6574\u9ad4\u4f86\u770b\uff0c\u9019\u4e0d\u662f\u8a13\u7df4\u65b0\u6a21\u578b\u7684\u9805\u76ee\uff0c\u800c\u662f\u4e00\u5957\u504f\u5411\u8a3a\u65b7\u8207\u89e3\u91cb\u7684\u5de5\u5177\u3002\u5b83\u7684\u5275\u65b0\u9ede\u5728\u65bc\u7528 minimal contrastive pairs \u52a0\u4e0a representation-level analysis\uff0c\u628a\u300c\u6a21\u578b\u70ba\u4f55\u6703\u7b54\u5c0d\u6216\u7b54\u932f\u300d\u62c6\u6210\u66f4\u5177\u9ad4\u7684\u5167\u90e8\u7d50\u69cb\u554f\u984c\uff0c\u5c0d\u60f3\u6df1\u5165\u7406\u89e3 VLM \u7a7a\u9593\u80fd\u529b\u7684\u4eba\uff0c\u50f9\u503c\u6bd4\u55ae\u770b\u6392\u884c\u699c\u66f4\u9ad8\u3002<\/p>\n<p><strong>GitHub\uff1a<\/strong> <a href=\"https:\/\/github.com\/cheolhong0916\/contrastive-probing\" rel=\"noopener noreferrer\">https:\/\/github.com\/cheolhong0916\/contrastive-probing<\/a><\/p>\n<p><strong>\u9805\u76ee\uff1a<\/strong> <a href=\"https:\/\/cheolhong0916.github.io\/whyfarlooksup.github.io\/\" rel=\"noopener noreferrer\">https:\/\/cheolhong0916.github.io\/whyfarlooksup.github.io\/<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u9019\u500b\u9805\u76ee\u4e0d\u662f\u518d\u505a\u4e00\u6b21\u7b54\u984c\u6e2c\u8a66\uff0c\u800c\u662f\u76f4\u63a5\u89c0\u5bdf VLM \u5167\u90e8\u600e\u6a23\u8868\u793a\u5de6\u53f3\u3001\u9ad8\u4f4e\u8207\u9060\u8fd1\u3002\u5c0d\u60f3\u7814\u7a76\u6a21\u578b\u504f\u5dee\u7684\u4eba\u5f88\u6709\u53c3\u8003\u50f9\u503c\u3002<\/p>\n","protected":false},"author":8,"featured_media":8624,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","wpai_meta_description":"","footnotes":""},"categories":[133,179,132,127,149,186,197],"tags":[],"class_list":["post-8625","post","type-post","status-publish","format-standard","hentry","category-133","category-nvidia","category-3d","category-127","category-149","category-186","category-framework"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8625","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=8625"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/8625\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/8624"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=8625"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=8625"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=8625"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}